Reproduce “Hyperbolic VAE via Latent Gaussian Distributions” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Seunghyuk Cho, Juyong Lee, Dongwoo Kim
Researchers from POSTECH and KAIST developed the Gaussian Manifold VAE (GM-VAE), which leverages the inherent hyperbolic geometry of Gaussian distributions to create a stable and efficient latent space. The GM-VAE consistently outperforms existing hyperbolic VAEs in density estimation and achieves robust state representation learning in model-based reinforcement learning.
Key insights
- The parameter space of univariate Gaussian distributions, equipped with the Fisher Information Metric, intrinsically forms a hyperbolic space.
- KL divergence locally approximates the squared Fisher-Rao distance, enabling its use to define a new latent distribution (PGM normal) with tractable analytical properties.
- Factorizing the PGM normal distribution into familiar Gaussian and Gamma components resolves issues of numerical instability and computational cost by enabling direct sampling and exact KL divergence computation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.